Skip to content

[OPTIMIZER] Fix MoveOpAfterLayoutConversion to bail out when there is a Select Op #2678

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Jokeren merged 4 commits into from
Nov 21, 2023
Merged

[OPTIMIZER] Fix MoveOpAfterLayoutConversion to bail out when there is a Select Op #2678

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ac985c8
Fix MoveOpAfterLayoutConversion to bail out when there is a Select Op
manman-ren Nov 18, 2023
a68986a
add test case
mren2 Nov 18, 2023
0844120
Instead of bailing out for select, use correct newCvtTy for each oper…
mren2 Nov 20, 2023
d7775b1
fix clang-format
mren2 Nov 20, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 10 additions & 5 deletions lib/Dialect/TritonGPU/Transforms/OptimizeDotOperands.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,8 @@ class MoveOpAfterLayoutConversion : public mlir::RewritePattern {
if (isa<triton::LoadOp>(currOp))
checkOp = true;
else if (checkOp) {
// Bail out if there exists an op after Load that is not FpToFp,
// Bitcast, or Arith.
if (!isa<triton::FpToFpOp, triton::BitcastOp>(currOp) &&
currOp->getDialect()->getTypeID() !=
mlir::TypeID::get<arith::ArithDialect>())
Expand All @@ -134,19 +136,22 @@ class MoveOpAfterLayoutConversion : public mlir::RewritePattern {
// only considers conversions to dot operand
if (!cvtTy.getEncoding().isa<triton::gpu::DotOperandEncodingAttr>())
return mlir::failure();
auto argTy = cvtArgOp->getOperand(0).getType().cast<RankedTensorType>();
auto retTy = cvtArgOp->getResult(0).getType().cast<RankedTensorType>();
if (!argTy || !retTy)
if (!retTy)
return mlir::failure();
Type newRetTy = RankedTensorType::get(
retTy.getShape(), retTy.getElementType(), cvtTy.getEncoding());
Type newCvtTy = RankedTensorType::get(
retTy.getShape(), argTy.getElementType(), cvtTy.getEncoding());
int numArgs = cvtArgOp->getNumOperands();
SmallVector<triton::gpu::ConvertLayoutOp> newCvts(numArgs);
for (int i = 0; i < numArgs; i++)
for (int i = 0; i < numArgs; i++) {
auto argTy = cvtArgOp->getOperand(i).getType().cast<RankedTensorType>();
if (!argTy)
return mlir::failure();
Type newCvtTy = RankedTensorType::get(
retTy.getShape(), argTy.getElementType(), cvtTy.getEncoding());
newCvts[i] = rewriter.create<triton::gpu::ConvertLayoutOp>(
cvt.getLoc(), newCvtTy, cvtArgOp->getOperand(i));
}
auto newRet = rewriter.clone(*cvtArgOp);
for (int i = 0; i < numArgs; i++)
newRet->setOperand(i, newCvts[i]);
Expand Down
21 changes: 21 additions & 0 deletions test/TritonGPU/dot-operands.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,3 +153,24 @@ tt.func @mma_v3_reg_operand_A(%arg0: tensor<128x64xf16, #mma>, %arg1: tensor<64x
tt.return %r : tensor<128x64xf32, #mma>
}
}

// -----
#blocked = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [4, 1], order = [0, 1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [0, 1]}>
#mma = #triton_gpu.mma<{versionMajor = 2, versionMinor = 0, warpsPerCTA = [2, 2], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0], instrShape = [16, 8]}>
module attributes {"triton_gpu.compute-capability" = 80 : i32, "triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : i32, "triton_gpu.threads-per-warp" = 32 : i32} {
tt.func @a_impl(%pa: tensor<128x128x!tt.ptr<f16, 1>, #blocked>) -> tensor<128x128xf32, #mma> {
%cst = arith.constant dense<0.000000e+00> : tensor<128x128xf32, #mma>
%cst_0 = arith.constant dense<0.000000e+00> : tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma, kWidth = 2}>>
%cst_3 = arith.constant dense<5> : tensor<128x1xi32, #blocked>
%cst_4 = arith.constant dense<0.000000e+00> : tensor<128x128xf16, #blocked>
%tl = tt.load %pa {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<128x128xf16, #blocked>
%tr = tt.make_range {end = 128 : i32, start = 0 : i32} : tensor<128xi32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>
%te = tt.expand_dims %tr {axis = 1 : i32} : (tensor<128xi32, #triton_gpu.slice<{dim = 1, parent = #blocked}>>) -> tensor<128x1xi32, #blocked>
%tc = arith.cmpi slt, %te, %cst_3 : tensor<128x1xi32, #blocked>
%tb = tt.broadcast %tc : (tensor<128x1xi1, #blocked>) -> tensor<128x128xi1, #blocked>
%ts = arith.select %tb, %tl, %cst_4 : tensor<128x128xi1, #blocked>, tensor<128x128xf16, #blocked>
%conv = triton_gpu.convert_layout %ts : (tensor<128x128xf16, #blocked>) -> tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma, kWidth = 2}>>
%td = tt.dot %cst_0, %conv, %cst {allowTF32 = true, maxNumImpreciseAcc = 0 : i32} : tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #mma, kWidth = 2}>> * tensor<128x128xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #mma, kWidth = 2}>> -> tensor<128x128xf32, #mma>
tt.return %td : tensor<128x128xf32, #mma>
}
}